KiqStat
A football prediction and betting-analysis platform built for the Nigerian market.
Real fixtures, real live scores, and a statistical model fitted on actual completed
matches — no placeholder games, no invented picks.
What this is
Most football tipster sites hand you a selection and ask you to trust it. This one
computes probabilities from a published method and shows you the sample size behind
every number, including when that sample is too thin to say anything at all.
- Live scores across every competition the configured feeds carry.
- Fixtures up to 14 days out, kickoff times in West Africa Time.
- Predictions for 1X2, over/under, both-teams-to-score, double chance and
correct score — all read off a single scoreline distribution, so they cannot
contradict one another.
- Value detection that removes the bookmaker's margin before comparing prices,
with capped Kelly staking guidance.
- Bet slip builder showing an accumulator's true combined probability, and
warning when legs are correlated or the slip has grown long enough that the
margin dominates.
- Model transparency on every match: fitted ratings, sample size, data quality
and the parameters used.
Quick start
npm install
npm run dev
Open http://localhost:3000. No API key is required — the app falls back to
TheSportsDB's public key, which returns real live scores and real fixtures. See
Data providers for lifting the coverage limits.
npm run build # production build
npm test # unit tests for the model, odds and staking maths
npm run lint
npm run smoke # hit the live feeds and print what comes back
Data providers
Every fixture, scoreline and result comes from a live provider. Nothing is
generated or seeded. Adapters are merged rather than ranked: the same fixture from
two sources is de-duplicated on the club pair and kickoff day, the richer record
wins, and gaps are filled from the other.
| Provider | Env var | Coverage | Free tier |
|---|
| football-data.org | FOOTBALL_DATA_API_KEY | Top European competitions, untruncated result sets. Recommended. | 10 req/min |
| API-Football | API_FOOTBALL_KEY | Widest coverage — the only one carrying NPFL and CAF properly | 100 req/day |
| TheSportsDB | THESPORTSDB_API_KEY | Used automatically with the public key when nothing else is set | Public key caps list endpoints at a few rows |
Copy .env.example to .env.local and fill in whichever you have. With none set,
the app runs on the shared public key and says so in the interface — live scores
stay complete, but fixture lists are capped upstream.
An optional ANTHROPIC_API_KEY upgrades the written match analysis from the
deterministic generator to Claude. Claude is given the computed numbers and asked
to explain them; it is never asked to predict anything, and the deterministic
version is used whenever the call fails.
How the model works
For a fixture between home side i and away side j:
λ = exp(intercept + attack_i − defence_j + homeAdvantage) home goals
μ = exp(intercept + attack_j − defence_i) away goals
Ratings are fitted by weighted maximum likelihood using a diagonal Newton step.
The log-likelihood of a Poisson model with a log link is concave, and the L2 penalty
makes it strictly concave, so the optimum is unique.
- Dixon-Coles correction. Pure Poisson understates 0-0, 1-0, 0-1 and 1-1,
because teams play differently when a game is tight. The dependency parameter ρ
is fitted alongside everything else.
- Recency weighting. Match weight halves roughly every six months, so current
form outranks last season.
- Shrinkage. An L2 penalty pulls thinly-observed clubs toward league average,
which stops a promoted side with three games played from getting an extreme
rating off one fluke result.
- Identifiability. The intercept is a fitted parameter with both rating
vectors centred. Pinning it instead forces the overall scoring level into the
penalised defence ratings, and the home advantage collapses toward zero to
compensate — there is a regression test for exactly this.
- Early season. When the current campaign has too few completed matches, the
loader walks back through previous seasons until the sample is usable, trying
both season-label conventions (
2026-2027 and 2026).
The honesty gate
A prediction is only published as a pick when the history behind it supports one.
Below 15 completed matches in the competition, or fewer than 3 appearances for
either club, the model output is still shown — with an explicit warning — and no
selection is offered. Fixtures in that state are listed separately on the
predictions page rather than quietly mixed in.
What it cannot see
The model knows results. It does not know about suspensions, injuries, midweek
fixture congestion or a waterlogged pitch. And a probability is not a forecast: a
65% home win will be wrong about one time in three, which is the model working
correctly rather than failing.
Architecture
src/
app/
(app)/ live, fixtures, predictions, match/[id], trends, slip, content pages
api/ live, fixtures, match/[id], trends, health
components/
landing/ hero, marquee, feature and pricing sections
match/ cards, market panels, value calculator, slip
ui/ design-system primitives
lib/
providers/ three adapters, resolver, TTL cache with request coalescing
model/ the prediction engine — fitting, inference, pricing
ai/ analyst with deterministic fallback
service.ts the seam the pages and API routes share
Provider responses are cached in-process with TTLs tuned to each feed's rate limit,
and concurrent requests for the same key are coalesced so a cold cache under load
produces one upstream call, not many.
Tests
npm test
Covers the scoreline distribution and its normalisation, the Dixon-Coles
correction, market coherence (1X2 partitions the sample space; over and under are
complementary at every line), recovery of known team strengths from simulated
seasons, the score equations at convergence, margin removal, Kelly capping and
accumulator maths.
Responsible gambling
Sports betting in Nigeria is regulated and restricted to adults 18 and over. This
project is an analytics tool: it does not take bets or hold funds. Predictions are
statistical estimates, not certainties.
Licence
Provided as-is. Match data belongs to the respective providers and is subject to
their terms.
